What Generative Engine Optimisation Can and Cannot Do
- Chris Welford
- Aug 10
- 15 min read
The short version
Generative engine optimisation is the practice of getting a business cited inside AI-generated answers rather than ranked in a list of links.
Ofcom found around 30% of UK searches now show an AI Overview, and ChatGPT took 1.8 billion UK visits in the first eight months of 2025.
Pew Research Centre found US click-through halved when an AI summary appeared, and only 1% of users clicked a link inside the summary.
Aggarwal et al. (KDD 2024) found content optimisation could raise visibility by up to 40%; keyword stuffing performed below baseline.
Puerto et al. (NeurIPS 2025) found most conversational SEO methods largely ineffective under competitive conditions, with gains shrinking as adoption rises.
Cloudflare measured crawl-to-refer ratios in June 2025 of roughly 14:1 for Google, 1,700:1 for OpenAI and 73,000:1 for Anthropic.
Blocking a training crawler such as GPTBot does not remove a site from that company's AI search answers, which a separate crawler handles.
There is no reliable AI equivalent of rank tracking; a fixed monthly prompt panel plus GA4 referral data is the practical substitute.
What Generative Engine Optimisation Can and Cannot Do
Generative engine optimisation, or GEO, is the practice of getting your business named, quoted, or cited in the answers that AI tools like ChatGPT, Google's AI Overviews, and Perplexity produce. It matters because those answers now sit between a customer's question and your website, and if the model never mentions you, a growing share of people never reach you at all.
I run a marketing agency in Liverpool, and this is the question I get asked more than any other. Owners have started noticing their own customers saying "I asked ChatGPT" where they used to say "I googled it". They are right to pay attention. What follows is what the evidence actually supports, including the parts that undercut the way most agencies are currently selling this. I would rather you read that from me than find it out after you have signed something.
What is generative engine optimisation?
The clearest definition comes from the academic paper that named the field. Aggarwal and colleagues at Princeton and Georgia Tech, presenting at the 30th ACM SIGKDD conference in Barcelona in August 2024, defined a generative engine as a system that uses generative models to gather and summarise information in order to answer a user's query. That is the important structural point. A traditional search engine retrieves documents and ranks them for you to choose between. A generative engine retrieves documents and then writes the answer itself, choosing on your behalf which sources to lean on and which to leave out.
GEO is the work of influencing that second decision. Traditional SEO tries to rank your page inside a list. GEO tries to get your business quoted, named or linked inside the written answer that increasingly appears above that list. The goal shifts from "rank on page one" to "be the source the model decides to trust".
That is a genuine change, but it is not a clean break. A generative engine still has to retrieve your page before it can summarise it, and retrieval is largely the same problem search has always been. Anyone telling you that SEO is dead and GEO replaces it is selling a story rather than describing a system. GEO sits on top of good SEO. It does not stand in for it.
Why does GEO matter now?
Two sets of numbers make the case, and I want to be precise about what each one does and does not show. The first is UK data from Ofcom, published in the Online Nation 2025 report on 10 December 2025. Google Search is used by 82% of UK adults and handles around 3 billion searches a month in this country. About 30% of those searches now display an AI Overview, and 53% of adults say they see these summaries often. Ofcom also recorded ChatGPT taking 1.8 billion UK visits in the first eight months of 2025, up from 368 million across the same period in 2024. That is not a rounding error. That is a roughly fivefold increase in a year, in a single market, on a single product.
The second set concerns behaviour, and here the best evidence is American. Pew Research Centre published an analysis on 22 July 2025 based on the real browsing activity of 900 US adults during March 2025, covering 68,879 unique Google searches. Where an AI summary appeared, users clicked through to a traditional search result in 8% of visits. Where no summary appeared, they clicked in 15% of visits. Users clicked a link inside the summary itself in just 1% of visits. Browsing sessions ended entirely on 26% of pages carrying an AI summary, against 16% of pages without one.
Read that 1% again. When the machine answers the question properly, most people stop. They have what they came for.
I want to flag the limitation honestly rather than paper over it, because I see this study quoted constantly with the caveats stripped out. Pew's sample is 900 US adults on Google specifically, in one month, over a year ago. It is not a UK figure; it does not cover ChatGPT or Perplexity, and search interfaces have changed since March 2025. What it establishes is a direction and a mechanism, not a number you can apply to your own traffic. The Ofcom data tells us the UK exposure is real and large. Pew tells us what tends to happen when people are exposed. Putting those together is reasonable. Pretending Pew measured British consumers is not.
How is GEO different from SEO?
The two disciplines share a great deal of DNA, but they reward different things.
Traditional SEO | Generative engine optimisation |
Goal: rank in the list of links | Goal: get named in the AI answer |
Success is a click to your site | Success is a mention, a citation or a click |
Rewards keywords and links | Rewards clear answers and trusted mentions |
Measured in rankings and traffic | Measured in citations and brand mentions |
One surface, the search engine | Many surfaces: ChatGPT, Perplexity, AI Overviews, Copilot |
Position is stable and checkable | Output varies between sessions and users |
That last row does more damage than people expect, and I will come back to it when we get to measurement.
The overlap is real. A quick site, clean information architecture and honest, well-organised content help you in both jobs. GEO simply adds a requirement: write and structure the content so a machine can lift a clean, correct, self-contained answer directly off the page without having to interpret you.
What does the research actually say works?
The Aggarwal paper tested nine content strategies against GEO-bench, a benchmark of roughly 10,000 queries spanning multiple domains. The headline finding, stated in the abstract, is that GEO methods can improve a source's visibility in generative engine responses by up to 40%. The strategies that carried most of that gain share a family resemblance. Adding quotations from credible sources. Adding statistics. Citing sources explicitly. Improving fluency, which is to say writing more clearly without adding any new information at all. What links them is that each one makes a passage easier to lift and attribute. A sentence containing a figure and a named source is a sentence a model can quote with confidence, because the provenance travels with the claim.
Two findings deserve more attention than they get. The first is that fluency optimisation produced meaningful gains without adding a single new fact, which suggests that dense, hedged, badly-structured prose actively works against you even when the underlying information is good. The second is that keyword stuffing performed below the unmodified baseline. Loading a page with repeated query terms made it less likely to be cited, not more. The tactic that defined a decade of bad SEO appears to be actively counterproductive here. There is a pleasing irony in the fact that the research recommends citing sources and quoting statistics, and that this article does exactly that. That is not accidental. This piece is a working demonstration of the advice, which you are welcome to check by asking an AI tool about GEO and seeing whether it turns up.
Where the evidence gets uncomfortable
Here is the part most agency blog posts leave out, and it is the reason I have written this section longer than the others. In June 2025, Puerto, Gubri, Green, Oh and Yun published C-SEO Bench, subtitled "Does Conversational SEO Work?", which was subsequently accepted at the NeurIPS 2025 Datasets and Benchmarks Track. They set out to test the GEO methods across a broader range of domains and, crucially, under realistic competitive conditions. Their finding is blunt. Most current methods are largely ineffective, contrary to results reported in earlier literature. Several actively hurt the document's ranking. Meanwhile, traditional SEO strategies, meaning those aimed at improving the source's ranking within the model's retrieved context, were significantly more effective. There is a second finding in that paper which I think is the most important sentence written about this field so far. As the number of actors adopting these techniques increases, the overall gains decrease. The authors describe the problem as congested and zero-sum. Sit with that. If the tactics work when you are the only one doing them, and stop working when everyone does them, then a large part of what is currently sold as GEO strategy has a shelf life measured against your competitors' adoption rate rather than against anything durable. This is the same arc SEO ran through between roughly 2005 and 2015, compressed into a much shorter period.
So how do we hold both papers at once? Honestly, I had a hard time understanding it until I realised that they are measuring different things under different conditions, and the disagreement is not yet resolved. Aggarwal measured single-source optimisation against a benchmark. Puerto measured competitive multi-actor adoption and included a straightforward retrieval-ranking baseline for comparison, which outperformed the clever content tricks by a wide margin in their retail testing. Both are peer-reviewed. Both are recent, as of 2025; I know in the AI universe it may as well be 100 years ago, but it's what we have to go off. Neither has been replicated enough times for anyone to claim the matter is settled.
What survives the disagreement is the boring part, and the boring part is the part I would build a strategy on. Both papers agree that a source which is genuinely relevant, retrievable and clearly written does well. They disagree about whether bolting extra quotations and statistics onto a mediocre page moves the needle once everyone is doing it. If your foundations are weak, no amount of quotation-stuffing rescues you, and the research suggesting otherwise was conducted in conditions that no longer describe the market.
This is why I will not sell a client a GEO retainer when their website loads in six seconds and their service pages read like a brochure from 2014. It would be taking money for the top floor of a building with no ground floor. That position has cost me work. I would rather it cost me work than cost a client eighteen months.
The crawler problem nobody puts in the proposal
There is a structural issue underneath all of this which affects whether GEO is even a sensible investment for a given business, and it comes from Cloudflare rather than academia.
Cloudflare publishes what it calls the crawl-to-referral ratio, how many pages a platform's crawlers request from sites, divided by the number of visitors that platform sends back. In June 2025, Cloudflare measured Google at roughly 14 crawls per referral. OpenAI came in at around 1,700 to 1. Anthropic sat at around 73,000 to 1. In a narrower sample covering 19 to 26 June 2025, Cloudflare Radar put Anthropic at 70,900 to 1 and Mistral at 0.1 to 1, meaning Mistral sent ten referrals for every page it crawled. Cloudflare's later analysis, published on 29 August 2025, found that training accounted for close to 80% of all AI crawling activity by the middle of that year.
The bargain that held for twenty years was simple. Crawlers took your content, search sent you visitors, visitors paid for the content. For a meaningful share of AI crawling, the second half of that bargain no longer applies. Your pages are being read at industrial scale to train systems that may never send anyone back.
The practical response is not to block everything in a fit of pique. It is to understand that the operators run separate crawlers for separate jobs, and that you can treat them separately. OpenAI runs GPTBot for training, OAI-SearchBot for ChatGPT's search product, and ChatGPT-User for live retrieval when someone asks a question. Google runs Googlebot for search and Google-Extended as the AI training control. Blocking a training crawler does not remove you from that company's search answers, because a different agent handles those.
So the sensible position for most small businesses is to allow the search and retrieval crawlers, since those are the ones that can actually cite you and send someone your way, and then make a deliberate decision about the training crawlers based on how much of your value sits in the content itself. A heating engineer's service pages and a publisher's archive are not the same asset, and they should not get the same robots.txt.
Check yours. Most of the small business sites I audit have never had that file looked at by a human being.
The measurement problem, which is worse than you think
Traditional SEO measurement works because search results are broadly stable and independently checkable. Two people in Liverpool searching the same term at the same time see close to the same page. You can track position, and position correlates with clicks in a way that has been studied for twenty years.
None of that transfers cleanly. Generative engines produce different outputs for the same question depending on session, phrasing, personalisation and model version. There is no position one. A citation is not a click, and Pew's finding that only 1% of users click links inside AI summaries means a citation is very often not a click. So you are measuring an outcome that is unstable, non-deterministic and only loosely connected to revenue.
What I would actually track, in order of usefulness:
Referral traffic from AI platforms in GA4, filtered by source. This is the only hard number in the list. It is small for most businesses today, and that is fine, because the trend matters more than the level.
A fixed prompt panel. Pick fifteen to twenty questions your customers genuinely ask, phrased the way they would phrase them, and run them against ChatGPT, Google and Perplexity on a set date each month. Record whether you appear, and record it consistently. It is crude, manual, and far better than nothing.
Branded search volume in Search Console. If AI answers are working for you, more people should be looking you up by name, having heard of you somewhere they did not click.
Server log evidence of which AI crawlers are reaching which pages, because a page no crawler has ever fetched cannot be cited by anything.
Be extremely careful of anyone quoting you an "AI visibility score" without explaining how it is computed and how much it varies between runs. I have seen the same brand score wildly differently on the same tool a week apart with no changes made to the site. If the number moves that much on its own, it cannot tell you whether your work moved it.
What can a business do this week?
Start with a baseline, because everything else is guesswork without one.
Ask ChatGPT, Google and Perplexity the questions your customers actually ask. Not "best marketing agency", which nobody types, but the real ones: "who can service a boiler in south Liverpool", "do I need a Gas Safe engineer to move a radiator", "how much should blinds cost for a bay window". Write down whether you appear at all, and which competitors do. That single hour tells you more than any audit tool.
Then fix the obvious. Put a plain, complete answer in the first two sentences of every key page, then explain underneath. Add an FAQ written in the words customers use rather than the words your industry uses. Make sure your Google Business Profile is complete and that its details match your website exactly, because inconsistency between them is a signal that you are not who you say you are. Add Organisation, LocalBusiness and FAQPage schema so machines can read your content in a format built for them. Check robots.txt is not blocking the crawlers you want.
Then leave it alone for a quarter, and rerun your prompt panel on the same date.
One caution about the mentions question. The research is consistent that references on trusted third-party sites help, and the temptation is to go and buy some. Do not. The directories and paid placements that are easy to buy are exactly the ones the models have learned to discount, and you will spend money teaching a machine that you are the sort of business that buys placements. Trade press, local news, a properly maintained Google Business Profile, and genuine customer reviews are slower, and they are what actually counts.
Who should not bother with this yet
I will say this plainly because most of my industry will not.
If your website is slow, if your service pages do not clearly state what you do and where you do it, if your Google Business Profile is half-finished, or if you are not currently converting the search traffic you already have, then GEO is not your problem. Your problem is upstream, and the research supports that reading: the strategy that held up best under competitive conditions in the C-SEO Bench testing was improving the source's underlying ranking, which is to say ordinary, unglamorous SEO done properly.
GEO earns its place once the foundations are solid and you are competing for visibility with businesses whose foundations are also solid. That describes a real and growing number of firms. It does not describe most of the small businesses currently being pitched AI visibility packages.
Nobody can guarantee you AI rankings. There is no position to rank in, the outputs change between sessions, and the two most serious academic studies in the field disagree about how much content tactics matter. Anyone promising you a guaranteed outcome in that environment either has not read the research or is hoping you have not. What I can tell you is that clear writing, honest sourcing, structured pages and a genuine reputation held up in both studies. That was also true before any of this arrived, which is either reassuring or annoying depending on how much you were hoping for a shortcut. If you want a hand working out where you actually stand in AI search today, that is the sort of thing we do most weeks at Constellation.
Book a free strategy call: https://calendly.com/constellationmarketingsolutions/30min
FAQ (for FAQPage schema, real-query phrasing)
What is generative engine optimisation (GEO)? GEO is the practice of getting your business named, quoted or cited inside the answers produced by AI tools such as ChatGPT, Google's AI Overviews and Perplexity. It differs from SEO in that the goal is being used as a source rather than ranking in a list of links.
Is GEO the same as SEO? No. SEO works to rank your page in the list of search results, while GEO works to get your business cited inside the AI-written answer above those results. They overlap heavily, and research published at NeurIPS in 2025 found that traditional SEO signals remain the most reliable route to being cited.
Does GEO actually work? The evidence is mixed, and the field is young. Aggarwal et al. (KDD 2024) found content optimisation could improve visibility by up to 40%. Puerto et al. (NeurIPS 2025) found most of those methods were largely ineffective under competitive multi-actor conditions, and that gains shrink as more businesses adopt them. Both agree that genuinely relevant, clearly written, retrievable content performs well.
How do you get your business cited by AI? Answer questions completely in the first two sentences of each page, structure content with clear headings and FAQs, earn mentions on trusted third-party sites, add Organisation, LocalBusiness and FAQPage schema, and make sure your robots.txt allows the AI search crawlers such as OAI-SearchBot.
Why does GEO matter in the UK in 2026? Ofcom's Online Nation 2025 report found that around 30% of UK searches now display an AI Overview, that 53% of adults see these summaries often, and that ChatGPT recorded 1.8 billion UK visits in the first eight months of 2025, up from 368 million a year earlier.
Do AI summaries reduce clicks to websites? Pew Research Centre's analysis of 68,879 US Google searches in March 2025 found users clicked a traditional result in 8% of visits where an AI summary appeared, against 15% where none did. Only 1% clicked a link inside the summary. The study covers US users on Google and should not be read as a UK figure.
Should I block AI crawlers? Not indiscriminately. Operators run separate crawlers for separate purposes: OpenAI uses GPTBot for training and OAI-SearchBot for its search product. Blocking a training crawler does not remove you from that company's AI search answers, so most businesses should allow search and retrieval crawlers while making a considered decision on training crawlers.
How do you measure AI visibility? There is no reliable equivalent of rank tracking. Practical measures are AI referral traffic in GA4, a fixed monthly panel of customer questions tested across ChatGPT, Google and Perplexity, branded search volume in Search Console, and server log evidence of AI crawler access.
Sources
Ofcom, Online Nation 2025, published 10 December 2025. Google Search used by 82% of UK adults, ~3 billion UK searches a month, ~30% of searches show AI Overviews, 53% of adults see them often, ChatGPT 1.8bn UK visits Jan–Aug 2025 vs 368m in 2024. https://www.ofcom.org.uk/media-use-and-attitudes/online-habits/from-apps-to-ai-search-how-the-uk-goes-online-in-2025
Pew Research Centre, "Google users are less likely to click on links when an AI summary appears in the results", 22 July 2025. 900 US adults, March 2025, 68,879 unique searches; 8% vs 15% CTR; 1% click inside summary; 26% vs 16% session end. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., Deshpande, A., "GEO: Generative Engine Optimisation", KDD '24: Proceedings of the 30th ACM SIGKDD Conference, Barcelona, 25–29 August 2024. DOI 10.1145/3637528.3671900. Preprint arXiv:2311.09735.
Puerto, H., Gubri, M., Green, T., Oh, S.J., Yun, S., "C-SEO Bench: Does Conversational SEO Work?", arXiv:2506.11097, June 2025; accepted NeurIPS 2025 Datasets and Benchmarks Track. https://openreview.net/forum?id=oTeixD3oZO
Cloudflare, "Control content use for AI training with Cloudflare's managed robots.txt", 1 July 2025. June 2025 crawl-to-refer ratios: Google ~14:1, OpenAI 1,700:1, Anthropic 73,000:1. https://blog.cloudflare.com/control-content-use-for-ai-training/
Cloudflare Radar, "The crawl before the fall of referrals", 1 July 2025. 19–26 June 2025 sample: Anthropic 70,900:1, Mistral 0.1:1. https://blog.cloudflare.com/ai-search-crawl-refer-ratio-on-radar/
Cloudflare, "The crawl-to-click gap", 29 August 2025. Training accounted for close to 80% of AI crawling by mid-2025.
